Six months that changed ChatGPT, Claude and Copilot (and what it means for your business)
Halfway through 2026, and the AI tools most businesses use every day have quietly become different products. New models, new pricing, and new rules about what happens to your data.
I've pulled together the changes from the first half of the year that actually affect how you work. Not the hype, not the benchmarks. The things I'd want to know if I were responsible for AI in a business right now, whether that business is a team of two or two thousand.
How much has actually changed since January? A lot.
A quick tally, because I don't think most people realise the pace.
ChatGPT has had three new model generations this year (GPT-5.3, 5.4 and 5.5), with a fourth already in preview. Six older models were retired in one go in February. In June, OpenAI rebuilt ChatGPT's memory completely, so it now learns your preferences and projects in the background instead of needing to be told what to remember. OpenAI is also planning to turn ChatGPT into a "super app" that bundles agents, coding tools and third-party services like Canva and Booking.com into one platform.
Claude has released five models since February: Sonnet 4.6, Opus 4.7, Opus 4.8, Fable 5 and Sonnet 5. One of those, Fable 5, needs a proper explanation, which you'll find below.
Copilot changed the most for businesses. Since April, the Copilot buttons inside Word, Excel, PowerPoint and OneNote only work with a paid Microsoft 365 Copilot licence. The app was redesigned in May, Claude became a selectable model, and Copilot Cowork went into general availability in June with a new billing model attached.
Here's the honest takeaway. If you or your team learned these tools six or twelve months ago, a good chunk of that knowledge is out of date, and there are almost certainly features you're paying for and not using. The gap between businesses that keep up and businesses that don't is getting wider every month.
Claude's new Fable 5 model: brilliant, but know the rules
Anthropic launched Claude Fable 5 in June. It's their most capable model yet, sitting in a new tier above Opus, and it's excellent for complex, long-running work. It also behaves differently from any AI tool your team has used before, and there are two things to understand before you put client or company data through it.
First, it can switch models on you. Fable 5 runs automated safety checks on every request. If a conversation touches cybersecurity or biology topics, it automatically switches to a different model, Opus 4.8, and the checks can flag perfectly normal content too. This happens whether or not the user chose it, and the checks scan everything in the conversation, including uploaded files and connected tools, so a flag can be triggered by content you didn't even type. There's a toggle in settings ("Switch models when a message is flagged") that pauses and asks you instead of switching silently. I'd recommend turning that on so the behaviour is visible rather than happening in the background.
Second, your conversations are kept. Prompts and outputs on Fable 5 are retained for 30 days for trust and safety purposes, and unlike other Claude models, Fable 5 doesn't support Zero Data Retention. Anything flagged by the safety systems can be kept for up to two years, and Anthropic staff can review flagged conversations, with access logged. Microsoft has restricted its own employees from using Fable 5 while its legal teams review the policy, which tells you how seriously large organisations are taking this.
For context on how new this territory is: Fable 5 was suspended worldwide by a US government export order three days after launch, and only returned on 1 July with stricter safety classifiers. Powerful tools, moving fast, rules still being written.
What to do: decide which data categories are appropriate to run through Fable 5, and put it in writing. If your AI policy doesn't yet cover model switching or retention differences between models, it's out of date. The other Claude models (Opus, Sonnet, Haiku) are unaffected by all of this.
Copilot: check your billing before you get a surprise
Two changes landed on 1 July.
Microsoft 365 base pricing went up. New pricing took effect across Business Basic, Business Standard, E3, E5 and F3 plans, so your all-in Copilot cost rises even if your Copilot licence itself hasn't changed. Business plans do gain an extra 50 GB of mailbox storage as part of it. If you're weighing up full Copilot, the promotional rate on Copilot Business has been extended through 30 September 2026, so the window to lock in the lower price is still open.
Cowork is now billed separately from your Copilot licence. This is the one catching people out. Copilot Cowork is the "do it for me" side of Copilot: you describe a task, it goes away and does the multi-step work across your apps. It reached general availability on 16 June, and while your Microsoft 365 Copilot licence still covers Copilot Chat and the in-app features, Cowork tasks now consume Copilot Credits billed on top.
What that means in money: roughly a penny per credit on pay-as-you-go, so a light task costs a pound or two and a heavy report-building task can run to several pounds each time.
And the catch: any organisation without usage-based billing configured in the Microsoft 365 admin centre by 1 July loses access to Cowork entirely. It's a hard cutoff, not a warning. If your team used Cowork during the free preview and it has suddenly stopped working, this is why.
Three practical notes:
Admins can set spending limits at tenant, group and individual user level, so enable it with guardrails rather than leaving it open
After any task finishes, typing /cost in the task window shows exactly how many credits it used, which is the fastest way to build a realistic budget
Model choice is a cost lever: Cowork currently runs on Anthropic's Opus 4.8 and Sonnet 4.6, and Microsoft is releasing a lower-cost model called Cowork 1 for everyday tasks
Two thirds of Google searches now end without a click
If your business relies on Google traffic, this number matters. New research from SparkToro and Similarweb found that 68.01% of US Google searches ended without a click in the first four months of 2026, up from 60.45% in 2024. Put differently, only 276 of every 1,000 searches now send a click to the open web.
AI answers at the top of the results are doing the job your blog posts used to do. What still works: branded searches, local business queries and high-intent "ready to buy" searches. Broad informational content is where the traffic is disappearing.
So if AI can answer the general "how do I" questions in your industry, your content needs to offer what AI summaries can't: proof, real examples, pricing clarity, tools, original data and your actual point of view. The strategic question is shifting from "are we ranking?" to "are we being cited?". Being the business AI mentions by name is the new being on page one.
Regulation watch
Three developments worth a line each, particularly if you have compliance or legal responsibilities:
EU AI Act: some high-risk compliance deadlines have been delayed, but the Article 50 transparency obligations, which require clear labelling of AI-generated content such as chatbots and deepfakes, still take effect on 2 August 2026. If you sell into or operate in the EU, check your customer-facing AI tools now.
Copyright and training data: a US court ruled against ROSS Intelligence for training its AI on copyrighted Westlaw content, rejecting the fair use defence. If you're building or commissioning custom AI, the provenance of training data is now a genuine legal risk, not a theoretical one.
Digital likeness: the US NO FAKES Act, targeting unauthorised AI replicas of real people's voices and likenesses, has cleared the Senate Judiciary Committee. Get written authorisation before using AI to replicate any real individual in marketing content.
Try this: the marginal gains audit (15 minutes)
Most AI conversations focus on overhauling whole processes. The quieter win is marginal gains: ten minutes shaved off a task your team repeats every day compounds into weeks of capacity over a year. This exercise finds those. It works in Claude, ChatGPT or Copilot.
List the eight to ten tasks you or your team repeat most often in a typical week (status updates, meeting prep, report formatting, inbox triage, chasing information)
Paste the list into your AI tool and ask: "For each task, tell me whether AI tools we already license could shorten it, roughly how much time per occurrence, and what the person would do differently"
Ask it to rank by annual hours returned, assuming realistic frequency
Ask: "Which single change has the best ratio of time saved to effort required to adopt?"
Pilot that one change for a fortnight, then measure
The point isn't transformation. It's that a 10% saving on the work you do every single day usually beats a 90% saving on something you do twice a year.
The bottom line
Six months, a dozen new models, two pricing changes and a new set of data rules. Most of it lands the same way whether you're a team of two or two thousand: the tools moved, the pricing moved, and the rules around your data moved with them.
The businesses getting the most from AI right now aren't the ones with the biggest budgets. They're the ones who've noticed the ground shifting and adjusted, one billing setting, one policy update and one workflow at a time.
If this has raised a question about whether your team's skills or your governance still match the tools you're running, I'd love to hear from you. You can get in touch here or sign up to the monthly newsletter here to get this roundup in your inbox.